See my other comment here [0] - I’ve not found it possible to actually have the agents follow the explicit instructions. Even single shot tasks once the context window fills up (which it does, very very fast, and repeatedly if you use subagents) will ignore very explicit instructions.
- discuss the plan, dos-and-donts, instructions and have them add it to a PLAN_FEATURE.md with the goal, non-goal, non-negotiables, technical details on the top.
- come up with phases to implement your thing. P1, P2, ... P8.. whatever
- in each phase add simple one line TODOs with as much explanantion of the task as needed e.g, - [ ] Update DB schema, - [ ] Generate migrations
- Ask the agent to go one phase or todo at a time depending on complexity.
- if the context fills up, /clear the window and just ask it to do the next line.
- if you wrote up all you care about in the plan properly, it'll follow what you need it to do.
[0] https://news.ycombinator.com/item?id=49036222